Docs
PETsARD YAML
Executor YAML (WIP)
Config Configuration

Config Configuration

Config is responsible for parsing and managing YAML configuration files, converting declarative configurations into executable module sequences.

Configuration Structure

YAML configuration files adopt a hierarchical structure, with each module containing one or more experiment configurations:

ModuleName:
  experiment_name:
    parameter1: value1
    parameter2: value2

Complete Configuration Example 

# Data loading
Loader:
  load_benchmark:
    filepath: benchmark://adult-income
  load_custom:
    filepath: ./data/custom.csv

# Data preprocessing
Preprocessor:
  preprocess:
    method: default

# Data splitting
Splitter:
  split_train_test:
    train_split_ratio: 0.8
    num_samples: 3

# Data synthesis
Synthesizer:
  sdv_gaussian:
    method: sdv
    model: GaussianCopula
  sdv_ctgan:
    method: sdv
    model: CTGAN

# Data evaluation
Evaluator:
  evaluate_quality:
    method: sdmetrics-qualityreport
  evaluate_privacy:
    method: anonymeter-singlingout

# Results reporting
Reporter:
  save_synthetic:
    method: save_data
    source: Synthesizer
  generate_report:
    method: save_report
    granularity: global

Module Execution Order

Config automatically arranges module execution in the following order:

  1. Loader - Data loading
  2. Preprocessor - Data preprocessing (optional)
  3. Splitter - Data splitting (optional)
  4. Synthesizer - Data synthesis
  5. Postprocessor - Data postprocessing (optional)
  6. Constrainer - Constraints (optional)
  7. Evaluator - Data evaluation (optional)
  8. Reporter - Results reporting (optional)

Experiment Naming Rules

Basic Rules

  • Experiment names must be unique within the same module
  • Can use letters, numbers, underscores, and hyphens
  • Cannot end with _[xxx] pattern (reserved for system internal use)

Valid Experiment Names 

Loader:
  load_data:           # ✓ Valid
    filepath: data.csv
  
  load-benchmark:      # ✓ Valid
    filepath: benchmark://adult-income
  
  load_custom_v2:      # ✓ Valid
    filepath: custom.csv

Invalid Experiment Names 

Loader:
  load_data_[1]:       # ✗ Invalid: Reserved pattern
    filepath: data.csv
  
  load_[test]:         # ✗ Invalid: Reserved pattern
    filepath: test.csv

Splitter Special Handling

When Splitter configuration includes num_samples > 1, Config automatically expands it into multiple experiments:

Original Configuration 

Splitter:
  split_data:
    train_split_ratio: 0.8
    num_samples: 3

After Automatic Expansion 

Splitter:
  split_data_[3-1]:
    train_split_ratio: 0.8
    num_samples: 1
  split_data_[3-2]:
    train_split_ratio: 0.8
    num_samples: 1
  split_data_[3-3]:
    train_split_ratio: 0.8
    num_samples: 1

This expansion process is automatic; users only need to specify the num_samples parameter in the configuration.

Multi-Experiment Configuration

Cartesian Product Execution

When multiple modules contain multiple experiments, Config generates cartesian product execution combinations:

Loader:
  load_v1:
    filepath: data_v1.csv
  load_v2:
    filepath: data_v2.csv

Synthesizer:
  method_a:
    method: sdv
    model: GaussianCopula
  method_b:
    method: sdv
    model: CTGAN

Execution combinations:

  1. load_v1 → method_a
  2. load_v1 → method_b
  3. load_v2 → method_a
  4. load_v2 → method_b

Result Naming

Results for each execution combination are stored with combination names:

Loader[load_v1]_Synthesizer[method_a]
Loader[load_v1]_Synthesizer[method_b]
Loader[load_v2]_Synthesizer[method_a]
Loader[load_v2]_Synthesizer[method_b]

Configuration Validation

Config performs the following validations during initialization:

1. Structure Validation

  • Check if configuration is a valid dictionary structure
  • Verify module names are correct

2. Naming Validation

  • Check if experiment names use reserved patterns
  • Verify experiment names are unique within modules

3. Parameter Validation

  • Parameter checking performed by each module’s Adapter
  • Validation occurs during Adapter instance creation

Error Handling

ConfigError

When configuration doesn’t meet rules, ConfigError is raised:

from petsard.exceptions import ConfigError

try:
    config = Config(config_dict)
except ConfigError as e:
    print(f"Configuration error: {e}")

Common errors:

  • Using reserved experiment naming patterns
  • Incorrect configuration structure
  • Missing required parameters

Notes

  • Execution Order: Config automatically arranges correct module execution order; manual specification unnecessary
  • Experiment Combinations: Multi-experiment configurations generate cartesian products; note execution time
  • Naming Conventions: Follow experiment naming rules to avoid using reserved patterns
  • Splitter Expansion: num_samples parameter automatically expands; manual configuration of multiple experiments unnecessary
  • Parameter Validation: Recommended to test configuration with small-scale data first, then run complete workflow after confirming parameters are correct
Status Tracking